PNW
Pacific Northwest
Research Station
F
I
N
D
I
N
G
I N S I D E
How Do We Tackle This Monster? . . . . . . . . . . . .
Teaming Gradients and Imputation . . . . . . . . . . . .
Objectives Under CLAMS. . . . . . . . . . . . . . . . . . . .
Evaluating Results Across the Landscape . . . . . . .
Landscape-Level Biodiversity Findings With GNN
Advantages of GNN for Integrated Assessments . .
Other Uses for GNN . . . . . . . . . . . . . . . . . . . . . . .
issue fifty-six / september 2003
S
“Science affects the way we think together.”
Lewis Thomas
SEEING THE TREES FOR THE FOREST: MAPPING
VEGETATION BIODIVERSITY IN COASTAL OREGON FORESTS
“Like the universal fascination
I N
with moving water, or the
dance of a fire’s flame, maps
S U M M A R Y
T
In order to address policy issues
relating to biodiversity, productivity,
and sustainability, we need detailed
understanding of forest vegetation at
broad geographic and time scales.
Most existing maps developed from
satellite imagery describe only
general characteristics of the upper
canopy. Detailed vegetation data are
available from regional grids of field
plots, but the data are not spatially
complete—they do not cover an
entire area of interest.
Prior to 1972, aerial photos were widely
used to map forest cover types across large
regions. But in the 1960s “in-place” (mapbased) inventories had been abandoned as
too expensive, and the plot-based approach
was adopted. Field plots, however, were only
ever as good as their geographic extent: no
fieldwork on earth could cover every inch of
ground. But with the increasing accessibility and decreasing cost of satellite imagery,
a vast new data resource was opened up.
“Every inch of ground” became a tempting
possibility.
Regardless of these limitations,
forest policymakers and stakeholders want information about current
forest conditions that is spatially
explicit (mappable), spans all ownerships, and is rich in detail, including
tree species, sizes, and densities.
Scientists studying regional vegetation patterns and dynamics require
similar data for their research.
hold some primal attraction
for the human animal.”
D. Aberley, Boundaries of Home
h e challenge of integrating Landsat
satellite imagery with ground-based
field plot data is unending. The process is replete with dead ends, loaded with
mathematical and statistical nightmares, and
likely to generate terminal eye rolls in the
most patient of researchers. The payoff, however, is mighty.
“As with any modeling approach, the design
was driven by the objectives of the users, and
by the constraints of the technology,” says
Credit: J.L Ohmann
The research and policy possibilities of
remote sensing were among the drivers of the
Coastal Landscape Analysis and Modeling
Study (CLAMS), a bioregional assessment
project focused on the Coast Range Province
of Oregon. Modeling vegetation by using
both remotely sensed data and field plot data
would be central.
Field plots, although expensive to establish
and maintain, are equally valuable components with satellite imagery of efforts to map
whole landscapes digitally.
Janet Ohmann, a research forest ecologist
with the Pacific Northwest Research Station
in Corvallis, Oregon. “In this case, we were
up against the limitations of satellite imagery. So if we’re interested in forest structure,
and the satellite is only giving us general
information about the forest canopy, we need
those field data more than ever to verify our
conclusions about more detailed characteristics of the forest.”
The Gradient Nearest Neighbor
method for mapping vegetation is a
breakthrough for regional assessments. Field plot, remotely sensed,
and environmental data are integrated into a single digital map. At
the regional scale, the method shows
an impressive level of accuracy,
suggesting its potential for strategic
regional planning across ownerships.
2
3
3
4
4
5
5
Of course, the CLAMS team was interested
in forest structure. The study’s wish list also
included forest species composition, age
classes, land use change, ownership, topography, and environmental gradients from the
coastal maritime forests to the dry inland
valleys. Every inch of ground. The quest for
a mapping approach became a process of
elimination, and lots of approaches didn’t
pan out.
“In the end we used components that were
not new individually, we just put them
together in a new way,” Ohmann explains.
For the database development and spatial analysis work of CLAMS, she teamed
up with Matt Gregory, an Oregon State
University faculty research assistant in the
Department of Forest Science. Several other
scientists on the interagency and interdisciplinary CLAMS team contributed valuable
ideas.
K EY FI N DI NGS
• The Gradient Nearest Neighbor (GNN) method of vegetation mapping integrates
field plot, remotely sensed, and environmental data into a single digital map.
GNN uses multivariate gradient analysis and imputation to assign a list of tree
species, densities, and sizes to each pixel in a regional map.
• GNN analyses revealed that variation in tree species composition was strongly
linked to physical environment and relatively unaffected by forest management
activities. Forest structure, defined by average tree size and density, was strongly associated with disturbance history and landownership patterns.
• Mid-successional, closed-canopy forests dominate the coastal landscape. Older
forest, along with legacies of snags and down dead wood, is concentrated on
federal lands; diverse young “natural” forest has been lost on all ownerships.
Foothill oak woodlands and hardwood trees are most abundant on nonindustrial
private lands.
• Evaluation at regional and site scales suggests that GNN maps are appropriately
used for regional-level planning, policy analysis, and research, not to guide local
management decisions.
HOW DO W E TACK LE T H IS MONST ER?
I
n case no one has noticed, the issues
involved in managing forests are growing alarmingly complex, or appropriately
complex, depending on your viewpoint. The
array of ecological and commodity values
and their interactions is daunting to any
manager, landowner, or interested citizen.
It is quite apparent that land use decisions,
environmental constraints, and ecological
processes constantly interact to influence the
spatial pattern of forested land cover.
“Therefore issues such as biodiversity conservation, long-term productivity and sustainability, and global climate change require
consideration of broad geographic scales,
from landscapes to whole regions, and long
time frames, from decades to centuries,”
Ohmann says.
Policy analysis considers the distribution
of forest resources and uses across multiple
ownerships, as well as landscape patterns
and forest conditions over time. Regional
assessments are starting to rely on simulation modeling to examine landscape change.
Vegetation maps are needed to support these
simulation models.
Most existing regional maps of forest cover
are based on classified satellite imagery,
Ohmann explains. “Although these data are
spatially complete, their information content
is limited to general characteristics of the
upper forest canopy. Few examples exist of
integrating this imagery with field plot and
environmental data for ecological modeling
and understanding at the regional scale.”
The Gradient Nearest Neighbor (GNN)
method Ohmann developed uses multivariate
analysis—dealing with many variables at the
same time—to assign a list of tree species,
densities, and sizes to each pixel in a regional map. For the first time in forest assessment, this allows us to integrate remote sensing, field data, and environmental data into a
single digital map. Breakthrough.
So OK, to achieve this you take tree lists
from field data, and include one tree list in
each mapped pixel? Given that a CLAMS
pixel is 25 meters on a side, that’s a boatload
of tree lists. How does this work?
Science Findings is online at:
Purpose o f
PNW Science F i n d i n g s
To provide scientific information
to people who make and influence
decisions about managing land.
PNW Science Findings is published
monthly by:
Pacific Northwest Research Station
USDA Forest Service
P.O. Box 3890
Portland, Oregon 97208
(503) 808-2137
Sherri Richardson Dodge, Editor
srichardsondodge@fs.fed.us
Send new subscriptions and
change of address information
to pnw_pnwpubs@fs.fed.us
Keith Routman, layout
kroutman@fs.fed.us
United States
Department of Agriculture
http://www.fs.fed.us/pnw
The site includes our new Science Update—scientific knowledge for pressing decisions
about controversial natural resource and environmental issues.
Forest Service
2
T EA M I NG GR A DI EN TS A N D I M PU TAT ION
A
technique called direct gradient
analysis is used to quantify relations
between vegetation and environment
for a sample of field plot locations. This
helps establish variations in tree species
composition and structure at different elevations and climates, topographic positions,
and disturbance histories.
Imputation is a statistical analysis tool for
incomplete data, whereby measured values
are assigned to sites lacking such
data. The GNN method, then,
refers to the practice of giving
a pixel the attributes of the plot
that is most likely to be similar in
terms of environmental and spectral characteristics. Think of it as
a best computer guess, after a lot
of input from researchers. A tree
list specifies what species, densities, and sizes of trees are likely
to occur at any given site, based
on ground measurements taken
on field plots. Tree lists can be
imputed to sites that have not been
sampled in the field.
Forest Inventory and Analysis (FIA) and
Old-Growth Study of the Pacific Northwest
Research Station.
Gradient analysis and imputation were the
existing tools that Ohmann and Gregory
put together in a new way for the CLAMS
project. They built their ideas on other
researchers’ efforts and questions, some of
which contributed, some of which did not
suit their needs. And they relied, of course,
on the growing computing power available
to them to come up with the digital parts of
the answers. Not surprisingly, computational
speed for simulation modeling is increasing
by orders of magnitude. Ohmann recalls that
a model run took 2 days at the beginning of
the CLAMS project 5 years ago. It now takes
2 hours, making a huge difference in how
quickly “What-if?” scenarios can be produced and evaluated, or sensitivity analysis
can be conducted.
“Direct gradient analysis and predictive vegetation mapping rest on
the premise that vegetation pattern
can be predicted from mapped
environmental data,” Ohmann
explains. “Predictive models are
based on various hypotheses as to
how environmental factors control
the distribution of species and
communities of trees and plants.”
For the CLAMS project, field
plot data were obtained from
the Current Vegetation Survey
(CVS) of the USDA Forest
Service and USDI Bureau of
Land Management; and the
Direct gradient analysis quantifies relations between vegetation and environment for a sample of
field plot locations. Imputation statistically assigns measured values to sites lacking such data. A tree
list specifies what species, densities, and sizes of trees are likely to occur at any given site, based on
ground measurements taken on field plots. Tree lists can be imputed to sites that have not been sampled
in the field.
OBJ ECT I V ES U N DER CLA MS
T
h e purpose of the GNN study within
the CLAMS project was to characterize, both quantitatively and spatially,
the current patterns of forest vegetation in the
Oregon coastal province. Under that general
rubric, specific objectives were threefold.
First, researchers wanted to quantify environmental and disturbance factors associated
with regional gradients of tree species composition and structure.
Second, they were to develop GIS-based
analytical tools and models to integrate field
plot, remotely sensed, and mapped environmental data to map current vegetation. And
third, they were to produce maps of current
vegetation that are model predictions. The
resulting vegetation map is used in landscape
simulations that project vegetation change as
far out in time as 100 years.
“The current vegetation map had to have sufficient fine-scale heterogeneity to support the
models of habitat capability for focal wildlife
and plant species being used in CLAMS,”
Ohmann says. “To be ecologically realistic,
we tried to develop a method of dealing with
multiple variables that would both predict
and maintain the integrity of different assemblages of tree species and structures.” Their
model also needed to interact with other
models—wildlife, aquatic habitat, and forest
dynamics.
In its mapped predictions, the study sought
to represent the full range of variability in
the area’s forest vegetation. Researchers also
worked on simultaneously mapping multiple
vegetation attributes that vary individually
and continuously, rather than separate, static
vegetation classes.
3
EVA LUAT I NG R ESU LTS ACROSS T H E LA N DSCA PE
T
h e predictive accuracy of the GNN
method was tested at the regional and
the site levels.
At the site level, a map of 10 vegetation classes defined by species composition, density,
and size class was 88 percent accurate within
one class. Classification accuracy improved
when fewer classes were used. Mapped predictions were most accurate for tree species
whose distributions are geographically limited
and strongly associated with climate, such as
Oregon white oak and Sitka spruce.
The ability to predict a given vegetation attribute with GNN, she points out, is limited by
the response variables used to develop the
underlying model. These response variables,
or summary measures, can be selected by the
researchers and tailored to study objectives.
“Improving prediction accuracy for some vegetation attributes may come at the cost of
reduced accuracy for others, and it is possible
to optimize the model for particular attributes,” she says. In other words, one model
Credit: R.J. Pabst
“The relative proportions of forest conditions
across the province predicted by GNN very
closely matched those estimated by systematic
grids of inventory plots,” Ohmann says. “This
agreement was not necessarily expected, even
though the GNN model was based on a subset
of the inventory plots. In addition, the mapped
GNN predictions reproduced the sampled
range of variability in vegetation across the
province very closely.” Furthermore, the GNN
approach does a reasonable job of portraying
fine-scale heterogeneity, Ohmann says.
Foothill oak woodlands and hardwood trees are most abundant on private nonindustrial
lands due to both environmental and past forest management effects. These forest types
are unprotected by existing forest policies and land use laws.
may best predict the location of recent
clearcuts or old-growth forest, but a different
model would be specified to predict the distributions of individual tree species.
“Similarly, perfect accuracy for multiple vegetation attributes in the GNN predictions is
impossible, because two plots never are
exactly alike nor are the vegetation and
explanatory factors perfectly correlated.”
It is important to note, she says, that there is
a danger that the fine spatial resolution and
detailed information content of the GNN
predictions may imply a higher level of precision than actually exists. “We stress that
vegetation maps produced with GNN are
appropriately used for strategic-level planning and policy analysis, not to guide local
management decisions.”
In time, she believes, as the technology
improves and other kinds of remote sensing
data become more affordable, the challenge
of accuracy at finer scales may gradually be
met.
LA N DSCA PE -LEV EL BIODI V ERSI T Y F I N DI NGS W I T H GN N
T
h e analyses generated by the GNNcreated maps revealed a variety of
new data about Oregon’s coastal
province. For example, variations in tree species composition are strongly linked to physical environment and relatively unaffected
(at least so far) by management activities.
In other words, Ohmann explains, tree species changed along a climate gradient from
the maritime coast to the inland valleys and
with elevation, but were little affected by
management activities. Therefore we see that
tree species still occupy essentially the same
ranges now as they did before timber harvesting began in the late 1800s.
Forest structure, on the other hand—defined
by tree size and density—was strongly associated with disturbance history, especially
human disturbance, as reflected in the satellite data and by land ownership patterns.
“The map revealed that mid-successional,
closed-canopy conifer forests dominated the
coastal landscape. Older forest was concentrated on federal lands, and over 90 percent
of foothill oak woodlands are on private
lands,” Ohmann observes.
A number of key biodiversity issues came to
light in the map, she notes. There has been
a loss of diverse young forest on all ownerships. Young stands in the current landscape
W R I T E R ’ S
have regenerated after clearcutting and other
intensive forest management activities and
typically lack the legacy of scattered large
live and dead trees that remain after natural disturbances such as wildfire and wind.
Conservation of diverse young forests has
received little attention in forest policy.
Large live trees, snags, and down dead wood
are present on all forest lands throughout
coastal Oregon, but are most abundant on
public lands. Many of these trees are legacy
from previously harvested or burned oldgrowth forest, and their presence in future
landscapes is uncertain. In contrast, foothill
oak woodlands and hardwood trees are most
P R O F I L E
Sally Duncan is a science communications analyst and writer specializing in natural resource issues. She is currently a Ph.D.
candidate in Environmental Sciences at Oregon State University in Corvallis.
4
abundant on private nonindustrial lands
owing to both environmental and past forest
management effects. Since these tree types
are underrepresented on public lands and in
reserves, they are unprotected by existing
forest policies and land use laws, Ohmann
notes.
“Our findings do suggest that some aspects
of biodiversity are affected more by physical
environment than by land management practices,” Ohmann says. “Regional conservation
planning for forest plant species therefore
needs to consider protection across broadscale environmental gradients such as climate
and elevation. Conservation efforts focused
on forest structure and thus wildlife habitat,
in contrast, need to consider landownership
and the effects of forest management practices.”
LAN D M ANAGEM ENT I M PLICATIONS
• Regional conservation for forest plant species needs to consider broad-scale
environmental gradients; efforts focused on forest structure and wildlife habitat
need to emphasize management practices and ownership patterns.
• Foothill oak woodlands are unprotected by existing forest policies and land use
laws, and are underrepresented on public lands and in reserves.
• Existing large live trees and dead wood are legacies from previously harvested
old-growth forests and have an uncertain future.
• Because the GNN method represents vegetation attributes in the map as individual continuous variables, user-defined classification schemes can be developed
and constructed for specific purposes.
A DVA N TAGES OF GN N FOR I N T EGR AT ED ASSESSM EN TS
U
n like many other mapping techniques, the integration of gradient
analysis and imputation gives the
GNN approach several distinct advantages in
bioregional assessment projects.
First, information content is high, because
each individual pixel contains a list of trees
by species, size, and density. “Because the
vegetation data are preserved at this most
basic level of detail, user-defined classification schemes can be applied, maps constructed, and accuracy assessed for specific
analytical purposes,” Ohmann points out.
Second, because of the imputation of a single
nearest-neighbor plot to each pixel, the
closely related variations among predicted
species and structures within map units are
ecologically realistic. Likewise, the range of
variability present in the sampled stands is
maintained in the mapped predictions. Thus
if the field plots are representative of the
entire regional landscape, the GNN procedure and maps will reflect the inherent variability of the region.
Although single-species models often will
yield better predictions than multispecies
models for the same species, it is also true
that methods that model the response of
single species lose information about the
co-occurrence of multiple species within
samples, Ohmann notes. The GNN approach,
she says, ensures that predicted plant communities are realistic assemblages of species
and structures.
Direct gradient analysis in itself contributes
to knowledge about regional ecological gradients, so the ongoing science endeavor is
enhanced. And finally, the current vegetation
model supports other models, such as wildlife, stand, and landscape dynamics, within
the CLAMS study.
OT H ER USES FOR GN N
A
lthough the GNN approach was developed specifically for the CLAMS
project, it appears that it could be successfully applied to other regions, so long as
a representative sample of georegistered field
plots and mapped spectral and environmental data are available. Already, biodiversity
analyses of the GNN-based maps have been
used to address criteria and indicators of
sustainability under the Montreal Process for
the Oregon Department of Forestry’s forestry
program for Oregon assessment.
In addition, a recently funded 3-year study is
applying GNN to mapping of forest vegetation and fuels in Washington, Oregon, and
California. Results of this research, Ohmann
says, will provide input to models that assess
fuel loadings and fire risk at the strategic
level for broad regions.
“The GNN maps also provide new opportunities for assessing the distribution of ecological and economic values across land ownerships and policy objectives, and potentially
for identifying new ways to accommodate
competing interests within a region,” she
says.
Under the umbrella of CLAMS, the maps are
used to evaluate various policy alternatives
that address key issues in forest management, across ownerships. Although the GNN
approach may not yet be perfectly accurate
at local scales, its value as a regional assessment tool surely returns the investment in
dead ends and eye rolls.
FOR FURTHER READING
Ohmann, J.L.; Gregory, M.J. 2002.
Predictive mapping of forest composition
and structure with direct gradient
analysis and nearest-neighbor imputation
in coastal Oregon, U.S.A. Canadian
Journal of Forest Resources. 32: 725–741.
Spies, T.A.; Reeves, G.H.; Burnett, K.M.
[and others]. 2002. Assessing the
ecological consequences of forest policies
in a multi-ownership province in Oregon.
In Liu, J.; Taylor, W.W.; eds. Integrating
landscape ecology into natural resource
management. Cambridge, United
Kingdom: Cambridge University Press.
Wimberly, M.C.; Ohmann, J.L. In review.
A multi-scale assessment of human and
environmental constraints on forest
landscape dynamics. Landscape Ecology.
5
F
I
N
D
I
N
G
PRSRT STD
US POSTAGE
PAID
PORTLAND OR
PERMIT N0 G-40
S
U.S. Department of Agriculture
Pacific Northwest Research Station
333 SW First Avenue
P.O. Box 3890
Portland, OR 97208-3890
Official Business
Penalty for Private Use, $300
SCIENTIST PROFILE
JANET OHMANN is
a research forest ecologist with the Ecosystem
Processes Program of
the Pacific Northwest
Research Station, USDA
Forest Service. She earned
B.A. and M.F. degrees
from Duke University
and a Ph.D. in forest ecology from Oregon State University. Before
coming to the Forestry Sciences Laboratory
in Corvallis, she worked several years for
the Forest Inventory and Analysis Program
in Portland, developing methods for the eco-
logical basis for extensive forest inventory
and analysis. Her current research focuses
on understanding broad-scale patterns and
dynamics of plant species and communities
and their relationships with environment and
disturbance in Pacific Northwest forests.
Ohmann can be reached at:
Pacific Northwest Research Station/USDA
Forest Service
Forestry Sciences Laboratory
3200 SW Jefferson Way
Corvallis, OR 97331-4401
Phone: (541) 750-7487
E-mail: johmann@fs.fed.us
COLLABORATORS
Matthew J. Gregory, Oregon State
University (OSU)
Coastal Landscape Analysis and
Modeling Study team (PNW
Research Station, OSU, PNW
Region, Oregon Dept. of Forestry)
Forest Inventory and Analysis
Program (PNW Research Station)
The U.S. Department of Agriculture (USDA) prohibits discrimination in all its programs and activities on the basis of race, color, national origin, gender, religion, age, disability, political beliefs, sexual orientation, or marital or family status. (Not all prohibited bases apply to all programs.) Persons with disabilities who require alternative means for communication of
program information (Braille, large print, audiotape, etc.) should contact USDA’s TARGET Center at (202) 720-2600 (voice and TDD). To file a complaint of discrimination, write USDA,
Director, Office of Civil Rights, Room 326-W, Whitten Building, 14th and Independence Avenue, SW, Washington, DC 20250-9410 or call (202) 720-5964 (voice and TDD). USDA is an
equal opportunity provider and employer.